\label{sec:discussion}
The intent of the DayFilter process is to provide building analysts a means of quickly filtering diurnal patterns in time series building performance data. The objectives were to provide a link between top-down and bottom-up analysis approaches, specifically AFDD and model calibration. We discuss the results of the application with consideration to these two contexts including limitations encountered. Additionally, we address the topic of data quality and its importance in this process.
For the implementation of AFDD, DayFilter provides a means of investigating a dataset at the early state of implementation. For example, qualitative, rule-based approaches benefit specifically as they require the setting of threshold parameters based on expert analysis. Despite their lack of sophistication, this category of AFDD is quite popular in implementation and many of the tools in the literature are based on this paradigm \cite{Ulickey:2010ut}. This fact is due to the capability of many building management systems to incorporate these techniques easily. In case study 1, the building operations staff was able to do a deeper analysis of the AHU tripping fault discords to set an alarm which would notify staff if the phenomenon occurred again. This type of rule-creation could occur automatically. One limitation in using this process for AFDD is the observation that anomalous behavior won’t alway manifest itself as a discord. More systematic or gradual failures could blend into the frequent motifs or eventually create patterns frequent enough to form their own motif.
We found that a key feature of the process was the ability to modify temporal granularity (through number of windows per day, \(W\)) or the magnitude or shape difference granularity (through alphabet size, \(A\)) to increase or decrease the proportion days tagged as discords as compared to motifs. The number of patterns created when selecting smaller windows or larger alphabets greatly increases, which can be seen as a disadvantage in terms of interpretability. The trade-off is that these settings gives more resolution to the process, thus creating more tightly grouped clusters and more effectively detecting discords. Modulating this feature enables focus on coarse high-level patterns or more sub-hourly phenomenon. We have tested these parameters according to the two case studies and provided suggestions for initial settings as applied to energy data in these contexts. However, a user of the process may decide to tune these parameters for further investigate a dataset for the implementation of certain AFDD approaches.
For model calibration, the DayFilter approach provides a few improvements in occupancy pattern and diversity factor detection as compared to conventional day-typing techniques \cite{Abushakra:2001us,Duarte:2013gy}. It was observed in the case studies that the process was able to effectively distinguish patterns between cooling and heating seasons, occupied and unoccupied phenomenon according to operation schedules and other such profiles. This differentiation was done in an automated way and is an improvement as compared to the day-typing techniques reviewed in the literature. The influence of anomalous days is not averaged into the patterns and thus a better representation of typical performance can be achieved. The output of the clustering step within the process can be used an input to diversity factor calculations outlined in the literature.
As with any data-driven approach, the usefulness of the process is only as good as the quality and amount of data available. A statistical approach like this doesn’t rely on physical knowledge of the building systems and is thus prone to error with inaccurate data. We prequalified the case studies in this paper based on how rigorous the data quality process for the sensor networks already is. The data from case study 1 was from a centralized chilled water plant which had undergone extensive accuracy tests such as a heat balance test and third party sensor accuracy verification. These checks are newly mandated by the local jurisdiction and are part of a steadily growing improvement in the quality and cost of sensor networks. Case study 2 also had a modern, well-calibrated sensor network and data acquisition system.